How do you perform multilabel classification that is also a multiclass problem?

Question

I have a data set in which each row of data belongs to certain classes/labels.

text	class1	class2	class3
text1	pos	neg	na
text2	na	neg	na
text3	na	neu	na
text4	pos	neg	neg
text5	neg	neg	na

There are basically 4 classes with 3 labels each (pos, neg, neu, na). I suppose this is both a multiclass and multilabel problem. How do I approach this? I am using the BinaryRelevance function from multisklearn but the result always returns 2 classes only (0 and 1). What is the correct way to do this?

Answer 1

From the description this is not a multilabel problem because:

• Each of the three "classes" (columns) must have a label. In a multilabel problem every class is optional.
• Every "class" (column) appears to have a specific purpose subdivided into 4 labels. In a regular multilabel problem the labels are exchangeable, e.g. a document can have topics "sports" and "society" but not "politics", all of these labels have no order and no specific role distinct from the others.

It seems that you simply have three regular independent multiclass problems:

• problem 1 = predict "class 1"
• problem 2 = predict "class 2"
• problem 3 = predict "class 3"

Note: the word "class" for the columns is confusing because these don't correspond to the regular concept of class.